Fire Safety Journal 67 (2014) 121134

Contents lists available at ScienceDirect

Fire Safety Journal

journal homepage: www.elsevier.com/locate/firesaf

Incorporation of decision, game, and Bayesian game theory

in an emergency evacuation exit decision model
Bryan L. Mesmer n, Christina L. Bloebaum
2271 Howe Hall, Iowa State University, Ames, IA 50011, USA

art ic l e i nf o

a b s t r a c t

Article history:
Received 3 November 2013
Received in revised form
24 April 2014
Accepted 11 May 2014
Available online 6 June 2014

This paper explores the use of utility and game theory to model egress decisions for exit choices found in
evacuations. These mathematically rigorous theories serve as a basis for individual exit decision making
that captures interactions between evacuees. The model presented in this paper is fundamentally
different from traditional evacuation simulators that capture the exit selection behaviour through simple
heuristics or objective functions. A utility function based on energy consumption of exit alternatives is
created that captures evacuee risk preferences and beliefs. Multiple game forms are created to allow for
trade-offs between model delity and computational complexity. These models range from Bayesian
games to simplied normal games. Multiple examples and validations are used to show that the decision
analysis model developed here captures natural human tendencies and characteristics. This enables
creation of a high delity exit decision model that simulates exit selection of evacuees.
& 2014 Elsevier Ltd. All rights reserved.

Keywords:
Evacuation
Game theory
Bayesian game theory
Decision analysis
Exit selection

1. Introduction
An egress scenario can be viewed as a series of decision making
processes by the individuals involved. Many decisions are made by
evacuees during an egress situation. Evacuees interpret information from multiple sources, which provides new knowledge on
which to base their decisions. This knowledge contributes to
evacuees beliefs on their exit alternatives. Evacuees determine
their best course of action based on their beliefs and preferences.
ESM [1], Exodus [2], SGEM [3] and SIMULEX [4] are well known
evacuation programs that use heuristics or simple objective functions associated with exit alternatives to determine an evacuee's
destination goal. While these methods attempt to capture the
decision making process, the key component of evacuee interactions is missing. Vacate [57] is a non-nodal based egress model
that uses driving forces to simulate the individual's wants and
needs. While the driving force approach captures some human to
human interactions, a method with a proven mathematical foundation for solving human interaction problems is still needed. This
paper presents a novel method to model the decisions evacuees
make concerning egress alternatives, with a focus on the effects of
evacuee interactions. The mathematically rigorous method of
this paper captures the exit selection of the individual through

Corresponding author. Tel.: 1 515 294 5666.

E-mail addresses: [email protected] (B.L. Mesmer),
[email protected] (C.L. Bloebaum).
http://dx.doi.org/10.1016/j.resaf.2014.05.010
0379-7112/& 2014 Elsevier Ltd. All rights reserved.

simulation of the decision analysis process rather than a heuristicbased simulation of the evacuee's behaviour.
Past research [8] has examined the use of utility theory [9,10],
game theory [10] and the velocity-obstacle method [11] in
maneuvering decision analysis. Maneuvering decision analysis
provides a mathematical representation for the decisions and
interactions involved in determining the optimum velocity magnitudes and directions for maneuvering in a crowded room.
Maneuvering decision analysis uses a separate decision process
from that of selecting the best exit alternative. Past research
focused on the movement decisions evacuees ought to make to
avoid obstacles and other individuals in getting to a specic exit. In
this paper, a method is presented to determine the decisions
evacuees ought to make to identify the best exit from a set of
alternative exits. While closely related, these papers offer the
needed mathematical techniques to simulate two separate decision processes inherent to evacuations. Validations presented in
past research [1214] and in this paper show the usefulness of
decision analysis in capturing the decision making process of
individuals in egress situations.
Both the structural environment and other individuals in that
environment are considerations when individuals decide between
egress alternatives. Current evacuation simulators incorporate
people in the environment as a variable in the alternative's
objective function. For example, if an evacuee is a follower
personality, then an alternative with a larger the crowd would
be more attractive than one with less. These simulators do not
take into account the interactions that occur between evacuees
during the actual decision making process, but only capture the

122

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

end effect. The method presented in this paper simulates the

decision making process that determine the evacuee's behaviour
in exit selectionnot the behaviours directly.
Decision theory and game theory [9,10,15] provide attractive
approaches to properly simulate the decision making process of
the evacuees exit alternative selections. The evacuee's environment dictates the type of game that is played. Decision theory is
used when there are no interactions occurring with other evacuees. In this case, the decision is based only on the evacuee's
beliefs about the alternatives and the evacuee's preference. This
situation can occur if no one else is in the vicinity to interact with.
Alternatively, game theory is used when the evacuees interactions
with others affect their alternative selection. For example, for two
evacuees in a room, each evacuee's decision may impact the
other's decision. In the language of game theory, they are playing
a game with one another. These two theories provide a mathematically rigorous basis for modelling the decision making processes people should make.
A limited amount of research has been conducted in using
decision or game theory principles to determine which egress
alternative a person should choose. A literary review [16] concludes that the incorporation of evacuee interactions in simulations is very rare. One study [16] focuses on a zero sum 2-player
game in which one of the game's players is the population and the
other player is an ambiguous capacity restricting entity. This
approximation for the evacuee interactions does not capture the
true complexity of individual player interactions. Hoogendoorn
and Bovy [17] have examined the formation of a utility function
and the use of differential games to examine the decisions
associated with crowd ow. Other studies [1821] focused on
the nature of crowd ow using game theory principles. These
studies incorporate an objective function that approximates the
evacuee's desires. In this paper, a value function that captures a
single preference of the decision-maker is used as an objective
function.
This paper presents the incorporation of game theory based
evacuee interactions in Vacate-GT (Game Theory), a decisionbased evacuation simulator. The interactions examined focus on
games played between evacuees during exit alternative selection.
This research is novel in modelling exit selection games where the
players are the individual evacuees and the utility functions are
based on a value function. The value function is formed from the
energy consumption associated with each egress alternative. The
preference is to minimize the anticipated energy consumption.
The games are played in certain time intervals, usually related to
the amount of time a player takes to complete a walking step. In
this paper, decision, Bayesian and normal games are used to model

different evacuee situations. These games represent the actual

decision making process more realistically than a pseudo-game
with a crowd playing against a restriction entity. Furthermore, the
use of different games allows for varying degrees of model delity.
As previously stated, the research presented in this paper
deviates from traditional large-scale evacuation simulators, in
which heuristics or simple objective functions are used to approximate the general crowd ow and evacuee exit decisions. This
paper models the decision behind an evacuee's choice of one exit
alternative over another. Instead of modelling behaviours directly,
this paper offers methods that model the decision process of the
individuals that result in behaviours. This is an important distinction. The research focuses on forming a high delity decision
model at the individual evacuee level. The delity of the decision
model varies according to the number of players (i.e. evacuees)
being modelled. The process of using decision and game theory to
determine the choices of each individual is currently more
computationally expensive than traditional simulators. However,
as computational power increases, the ability to use decisionbased higher delity models on a large-scale becomes less
burdensome.

2. Environment and egress alternatives

In this study, an entire oor of a building is used as the test
environment, with multiple rooms, hallways, etc., forming sections
within the environment. The passable barriers between the section and non-section areas (e.g. stairways, doors, windows or any
area not encapsulated by the section) are considered exits. Egress
alternatives are any invisible or visible barrier that, when passed,
allows the evacuee to see into a region that was previously visually
obstructed. In this denition, an egress alternative is not limited to
a section exit. For example, in a section with multiple rooms, an
egress alternative may be a door that leads to another room in a
section (i.e. a door between rooms of the section). Current
alternatives are egress alternatives that an individual is presently
deciding upon due to their current location.
Fig. 1A shows an evacuee (represented by an X) with two
current egress alternatives in the black outlined section. The
decision region (represented by the checker pattern) is the area
made of walls and alternatives surrounding the evacuee. The
current egress alternatives are represented by blue dotted lines
while the exits for the section are represented by red dashed lines.
The blue dotted line egress alternatives of Fig. 1A represent the
barriers which, if passed, would allow for the individual to view a
different area of the section. Fig. 1B shows the egress alternatives

Fig. 1. Examples of decision regions and egress alternatives. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this
article.)

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

and the decision region for a different location in the section, with
a section exit as one of the egress alternatives. Fig. 1C shows two
egress alternatives in blue. The X individual may be able to see into
the regions in which individuals Y and Z reside. However, Y must
pass the blue barrier to see more of the region below and Z must
pass the other blue barrier to see more of the region to the left.
Establishing these meaningful decision regions are important
since the models developed here incorporate interactions with
other individuals. Additionally, these limit the number of alternatives any one individual might have and reduces the number of
individuals interacting directly with one another.

3. Egress alternative energy consumption

The foundational principle of decision and game theory is the
determination of each individual's expected utility associated with
each alternative. The expected utility that an evacuee designates
for each alternative is dependent on his or her individual preferences. This paper investigates an individual's preference of wanting to expend the least amount of energy while evacuating. The
expected utility of an evacuee's egress alternative is then a
function of energy. This is partitioned into the energy consumed
getting to the alternative and the energy consumed after passing
through the alternative to reach the section exit.

3.1. Energy consumed in moving

Much work has been performed on representing energy consumed in human movement [2227]. The equations for walking
and running that are used in this study are energy equations [28]
that are a function of a person's weight and velocity. For this paper,
the energy equations are in units of nutritional calories per second.
The Vacate-GT simulator does not determine the movement speed
of an individual through heuristics or the speed of the simulator's
population. The evacuee's velocity is determined through decision
analysis. The evacuee chooses between running and walking
depending on which mode of movement consumes the least
amount of energy to evacuate.
A notion that is concluded multiple times in the literature
concerning energy expenditure of human movement [2224,26,28]
is that a person's natural movement speed is the speed associated
with the ratio of minimum energy expended to distance. This ratio can
be determined by dividing the metabolic rate of walking by velocity.
The velocity that results in the minimization of this equation is
1.124 m/s, which is independent of weight. This velocity is the most
efcient speed for a person if their energy function rate is represented
by the exact metabolic rates presented in the literature [28]. What the
chosen metabolic rate equations do not consider are other factors that
affect movement such as a person's sex and age. Therefore, the
metabolic rate equations are modied in this paper to better reect
the natural walking speeds of the population.
While a normal distribution is used to describe the distribution
of human natural walking speeds in statistics [29], a triangular
distribution is used in this paper. The triangular distribution offers
an approximation for the normal distribution [30] while enabling
easier calculations due to a nite range and linear segments. This
trade-off in statistical delity is performed in order to increase
computational efciency. A person's most efcient walking speed is
randomly chosen based on their age and sex, using the triangular
inverse cumulative distribution function approximation [30]. A unique
energy movement function is then created for each individual based
on that most efcient walking speed. The metabolic rate equations are
modied by using a normalized input velocity, found using Eq. (1).
Similar methods are used to determine an evacuee's shoulder width

123

[31], height, and weight [32,33].

n 

1:124
eff

where, n is the normalized input individual's velocity (m/s); is the

individuals velocity (m/s); and eff is the most efcient individuals
velocity (m/s).
In extreme cases, evacuees may feel the need to move at their
maximum speeds. Limitations on the velocities of the individuals
are imposed and correlate to the most efcient walking speeds.
In order to create a realistic limitation, the historic barrier of 10 s
for the 100 m sprint is used as the maximum velocity for the top
performing young males. Using this velocity of 10 m/s, and
assuming that the men capable of these speeds are at the top
percentile of most efcient walking speeds, the relationship seen
in Eq. (2) is derived for use in determination of individual
maximum velocity.
max eff  10=2:16917 eff  4:61

where, max is the individuals maximum velocity (m/s).

3.2. Energy consumed getting to egress alternative

Energy that is expended getting to an egress alternative
consists of 5 energy portions: (1) the energy spent moving to
the alternative; (2) waiting to go through the alternative; (3) passing through the alternative; (4) due to the dangers of the section;
and (5) the pseudo-energy gained by following other people. The
energy that is spent moving to the alternative is simply the energy
expended in moving at a certain speed for a specic distance for
an individual, calculated from the metabolic energy equations
[28]. The distance used is the shortest distance from the individual
to the alternative. This movement energy is referred to as Emove.
This distance does not take into account the extra distance needed
to maneuver around obstacles or other individuals. The energies
associated with such intricate maneuvers is handled in the
velocity decision process, which is addressed in past research [8].
Each evacuee holds a believed energy loss for traveling through
an alternative. For example, a non-doorway alternative may have
no energy loss, while passing through a door or climbing over a
barrier may have high energy losses. The believed energy expense
to pass through an alternative is represented by Epass. Energy must
be spent waiting to pass through an alternative if a queue of
people has formed ahead of the individual. Each evacuee holds a
believed ow rate for each alternative. This ow rate is the
believed amount of people that can pass through the alternative
per second. This paper uses an equation based on population
density and speed [34] to update the evacuees ow rate beliefs.
The time that a specic evacuee (the person of interest, or POI)
will have to wait to pass through the alternative is dependent on
the other individuals arrival times. Since the decision regions are
rarely fully saturated with evacuees, the function to determine the
waiting time due to other people passing through the alternative
is discontinuous. The waiting time does not linearly decrease, but
may increase or decrease sharply depending on the POIs arrival
time relative to the arrival times of the other evacuees. The
owchart in Fig. 2 is used to determine the waiting time,
dependent on a particular velocity of the POI. In this gure, T is
arrival time (s); FR is ow rate (individuals/s); j is the index of the
individuals; i is the person of interest; nw is the next time segment
in which no waiting is required; and a is the alternative. The arrival
times are compared to those of the other people choosing the
alternative and if they are signicantly close (o 0.1 s), the wait
time for each of those individuals arriving at the same time is the
average of their wait times. The energy expended in waiting, Ewait,

124

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

where, Egetting is the energy consumed getting to the initial egress

alternative (nutritional calories); Emove is the energy consumed
moving (nutritional calories); Epass is the energy consumed passing
through the alternative (nutritional calories); Ewait is the energy
consumed waiting (nutritional calories); Eharm is the energy
consumed due to harm inicted by the environment (nutritional
calories); Efollow is the energy consumed following other individuals (nutritional calories); i is a person of interest; and a is the
egress alternative.
3.3. Energy consumed after egress alternative to section exit

Fig. 2. Individual waiting time owchart.

is calculated as the product of the time spent waiting and the

metabolic rate of standing still for the individual.
A general section hazard rate (i.e. the energy drain on individuals per second in a section) is used in order to represent the
harm enacted on an individual due to a hazardous environment.
The harm energy, Eharm, is calculated by the time spent in the
section (both in moving and waiting) multiplied by the hazard rate
of the section. A difcult human trait to capture in simulations is
the behaviour of individuals who prefer following people, those
who avoid people, and any mixture in between. In order to
represent this behaviour in egress alternative selection, a follow
rate is created that represents the theoretical energy gained or lost
by the individual per person that passes through an alternative.
For example, a positive follow rate would relate to an individual
that prefers to avoid others. The follow energy, Efollow, is calculated
as the product of the number of people choosing an alternative
and that have been seen passing through the alternative, and the
follow rate.
The total energy expended in getting to an alternative for a POI
is a function of the POIs beliefs on the other individuals in the
decision region (which will be addressed through game theory
formulations), beliefs on the environment, physical traits, and the
velocity. This is captured in the previous energy forms and
combined in Eq. (3).
Egetting;a;i Emove;a;i Epass;a;i Ewait;a;i;  i Eharm;a;i;  i Efollow;a;i;  i
3

The energy that is expended after passing through the chosen

alternative is equivalent to the energy spent reaching an exit of the
section. This energy expenditure consists of the energy spent
moving to a section exit, moving through egress alternatives,
waiting in line for egress alternatives, and the harm done by being
in the section. The route from the decision region egress alternative to the section exit is very important in the determination of
energy expended. The route is a set of all of the alternatives that
must be passed in order to reach a section exit. The energy spent
moving to a section exit is dened as the metabolic rate of
traveling at the evacuees most efcient velocity for the time it
takes to travel the distance of the route. While further studies
could be performed to determine the optimal velocity after the
decision region, this paper assumes the individual chooses his
most efcient velocity.
The energy spent passing through egress alternatives along the
route is simply the summation of the believed pass through
energies of the routes alternatives, with the exception of the
initial alternative. The energy spent waiting in line for the
alternatives along the route to the section exit is the product of
the metabolic rate of standing still and the summation of the
believed waiting times for the routes alternatives (not counting
the initial alternative). The energy consumed due to the harm of
the environment after the decision region egress alternative is the
product of the hazard rate and summation of the time spent
traveling the route and waiting along the route. The total energy
expended after passing through the chosen decision region egress
alternative for the POI is a function of the route, the POIs
environment beliefs, the POIs physical traits, and the POIs most
efcient velocity. This energy expenditure is represented in Eq. (4).
Eafter;a;R;i Emove;R;i Epass;R;i Ewait;R;i; Eharm;R;i

where, Eafter is the energy consumed in egressing the environment

after passing through the initial alternative (nutritional calories)
and R is the route (sequence of egress alternatives).

4. Beliefs
We can see from the energy expenditure equation formulation
that the beliefs held by an individual have an important role in
determining their best egress alternative. The POIs beliefs used here
include beliefs held on other individuals and beliefs held on the
environment. The beliefs held on individuals (by a POI) pertain to the
most efcient velocities of the other individuals (the opponents).
These beliefs are represented as probability distribution functions for
each opponent in the decision region. The POI must estimate how
much each opponent values each egress alternative in order to be
able to make decisions involving opponent interactions. Therefore,
the POI must estimate the opponents expected utilities. In order to
simplify calculations of the opponents egress alternative expected
utilities, the beliefs of the opponents most efcient velocities are
based off of the POIs most efcient velocity for the normal game
theory form. The opponents age and gender are taken into

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

consideration in this calculation. The beliefs held by the POI regarding the opponents most efcient velocities are determined using the
previously explained method to randomly determine an individuals
most efcient velocity. In this situation, the random variable in the
calculation is equivalent to the percentile of the POIs most efcient
velocity. Hence, the POI assumes the opponents are of the same
physicality as he is, but adjusts for differences in age and sex.
The beliefs held by an individual on the environment include:
whether a direct path exists between two alternatives; the
distance between alternatives that have direct paths between
them; the energy needed to pass through each alternative; the
waiting time of each alternative; the ow rate for each alternative;
and the section hazard rates. These beliefs may be singular values
or probability distributions. For ease of calculations, the belief of
the existence of a direct path between two alternatives is a single
Boolean value. Also, for simplicity, the energy needed to pass
through an alternative and the ow rate of an alternative are a
singular value. The beliefs on the distance between two alternatives and the waiting time for each alternative are represented
by probability distributions. These beliefs are unique to each
individual.
4.1. Belief updating
In the course of the evacuation, the evacuees are constantly
receiving new information. The new information is either in the
form of a probability distribution or of a singular value. If the new
information affects the beliefs of the individual, then these beliefs
must be updated. If the new information is of a singular value (as
is the case in the belief of whether a direct path exists between
two alternatives) then the prior beliefs are overwritten by the new
information. If the new information is in the form of a probability
distribution (as in the case for information concerning distance
between two alternatives and the waiting time for each alternative), then the new information is combined with the prior
information to form a new posterior distribution. The mathematical technique for combining probability distribution beliefs used
in this paper is Bayes theorem [35].

5. Decision theory method

A decision theory model is used in the situation where the POI
is the only individual in the decision region. In this model, the
egress alternative selection is determined as a function of the
individuals beliefs. In accordance with decision theory, the individual ranks the egress alternatives and chooses the most preferred one based on his preference. In order to rank the
alternatives, a value must be assigned to each one. A mathematical
method for dening a persons preferences numerically is utility
theory [9,10]. The utility function converts a value function (in this
case energy measurement) into a comparable singular utility for
the individual. In utility theory, a persons risk preference plays a
large role on utility determination. The utility function is formed
in such a way as to reect a persons risk preference. For a neutral
preference, a linear relationship exists between the energy value
and the utility. For a risk averse preference, the hyperbolic
absolute risk aversion (HARA) utility function is used [36]. The
risk aversion utility function can also represent a risk proverse
preference when the coefcient of absolute risk aversion is
negative in the HARA function. The evacuees coefcient of
absolute risk aversion is randomly selected from a general population risk distribution. The utility function used in this paper is
represented in Eq. (5), which is a commonly used function.

uE  1 c  e  cE
5

125

where, u( ) is the utility function, E is the energy consumed

(nutritional calories); and c is the risk aversion coefcient.
5.1. Utility in certain environment
An individuals beliefs are comprised of singular values when
the individual is certain about the environment. If the individual
has singular beliefs, the utility for each egress alternative is
determined by Eq. (6).
ua;b;R u  Egetting;a Eafter;a;R

where, b is the individuals believed environment. Egetting,a and

Eafter,a,R are calculated according to Eqs. (3) and (4), respectively.
Two factors that have not been discussed are how to determine
the individuals route and velocity. This is addressed in the next
section.
5.2. Route and velocity determination in certain environments
An individuals route is dened as the path between the rst
egress alternative passed and a section exit. The route is determined using the Dijkstra algorithm [37], with terminology and
problem framework for the algorithm described in [38]. Using this
method, the vertices are dened as the egress alternatives of the
section. The distance between two vertices is dened as the
energy spent moving between the two vertices (see Eq. (4), where
the route consists of these two vertices). As this equation is for a
certain environment, there is only a singular belief on the distance
between the alternatives and the waiting time of the second
alternative. The starting vertex in the Dijkstra Algorithm is the
current egress alternative in the decision region that is being
considered. The ending vertices are any egress alternatives that are
exits of the section. The Dijkstra algorithm is then performed,
based on these parameters. The resulting route is the one that
spans from the starting vertex (considered current egress alternative) to an ending vertex (section exit) and has the least energy
consumption.
The individuals velocity in reaching the initial egress alternative is determined according to the minimum energy consumed
in getting to the alternative. For each alternative in the decision
region, a particle swarm optimization (PSO) [39] search is performed to determine the optimal velocity magnitude. Once the
optimal route and velocity for each alternative in the decision
region are determined, the utility of each egress alternative is
calculated using Eq. (6). The individual chooses the alternative
with the greatest utility.

5.3. Utility in uncertain environments

The utility theory technique of expected utilities is used in
order to create comparable worth of egress alternatives in uncertain environments. The expected utility takes into account the
beliefs of the individual. A set containing all of the possible
environments believed by the individual is created. For example,
the individual may believe that a hallway might be 25 m long, as
well as 10 m long. These result in two different environments in
which beliefs are held. This set consists of all of the possible
combinations of time delays of each alternative and the distances
between alternatives. The expected utility for each egress alternative is then calculated using Eq. (7). In this equation, the
probability that the individual believes an environment is true is
multiplied by the utility of the egress alternative given that the
environment is certain. The formulation determining the best
route is performed for each possible environment. Once the
expected utility for each egress alternative is calculated, they are

126

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

compared. The alternative with the highest expected utility is the

best choice for the individual.
ua;i allbinB probBb;i  ua;b;R;i  Egetting;a;i  Eafter;a;R;i 

where, is expected; b is the individuals believed environment;

and B is the individuals believed environment set.
5.4. Examples
An example scenario is created in order to clearly represent the
decision making process of an individual when they are in a
decision region alone. The example is also used to show how the
different energy consumptions affect their decisions. The effects of
distance beliefs on the expected utilities of egress alternatives for a
risk neutral player are examined rst. The environment examined
is shown in Fig. 3. The individual in this example (represented by
the X) is a 40 year old male weighing 86 kg, with a most efcient
velocity of 1.56 m/s and a belief that the hazard rate of the section
is zero. In this layout the individual is equidistant from egress
alternatives 1 and 2. The egress alternatives are represented by
blue dotted lines. The distance between the section exit 0
(represented by the red dashed line) and alternatives 1 and 2 is
also equidistant. The individual is certain of the parameters (delay
time and energy to pass through) of egress alternatives 1 and 2, as
they are visible. What is unknown to the individual is the distance
between those alternatives and the section exit, as well as the
parameters of alternative 0 since it is visually obstructed by a wall.
The parameters of alternative 0 (such as the believed delay time
and energy to pass through), are the same in both alternative
1 and 2s expected utility calculation. The primary difference
between the two expected utilities will be due to the believed
distance between each alternative and the exit (alternative 0).
If both distance beliefs between alternatives 1 and 2, and alternative 0 are equivalent, then each alternatives expected utility is
identical. Hence, the individual is indifferent between the two
alternatives.
Different expected utilities for the alternatives are anticipated if
the distance beliefs of the individual between alternatives 1 and
2 and the exit are different. For example, assume that the distance
beliefs seen in Fig. 4 are believed. In this situation, a player with a risk
neutral preference would have an expected utility for alternative 1 to
be 1.639 and for alternative 2 to be  1.57. The individual would
choose alternative 2 given these expected utilities.
The effects of risk on decision making are also examined. Fig. 5
shows a distance belief related to the environment in Fig. 3.
In Fig. 5 the individual is more certain of his distance belief of
connection 02 than connection 01. An individual who is risk
neutral is indifferent to the alternatives as the expected distances
are the same, as are the expected utilities. Alternatively, for a risk
proverse individual, there is a chance connection 01 is shorter
than 02 and the individual is willing to risk the chance that it is

Fig. 4. Non-equivalent expected outcome beliefs.

Fig. 5. Equivalent expected outcome beliefs.

longer. If his risk coefcient is  0.06 (indicating a risk proverse

preference), then he chooses alternative 1 with an expected utility
of 15.0922 compared to the expected utility of alternative 2 at
15.0909. While these expected utilities appear to be very close in
value, the differences are actually quite important. Many factors
contribute to these expected utilities being close in value, such as
the relatively small absolute degree of risk aversion, a small
environment, and the identical expected outcomes of the beliefs.
This example is used to illustrate the impact of risk preferences on
individual decisions, where different preferences result in a different
rank ordering of the alternatives based on their mathematically
calculated expected utilities. Similar to the calculations of the risk
proverse individuals decision determination, a risk averse individual
chooses alternative 2. Related examples were performed to conrm
effects of different beliefs of the individual on decision making and of
incorrect beliefs being updated to correct beliefs (using Bayesian
updating), impacting the decisions made.

6. Game theory model

Fig. 3. Decision theory example scenario.

When multiple individuals are present in a decision region the

decisions made by an individual may affect the decision made by
another. These interactions are represented mathematically by an

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

individuals decision having an impact on another individuals

egress alternatives expected utility. Game theory [10,15] is used to
represent such interactions and to determine the decisions the
individuals ought to make [40]. In the games formed to represent
these interactions, the parameters that need to be dened are the
players, the players strategies, and the preferences of the players.
6.1. Players, strategies and preferences
Egress alternative selection games are played between individuals that are in the same decision region. In the egress alternative
game that is created in this paper, the players can choose which
egress alternative in the decision region to move towards and the
velocity to move. Therefore, the strategies of the game are
comprised of an alternative and a velocity. For example, an
individual with two possible egress alternatives examining four
possible velocities has the following strategy set: [(a1,V1), (a1,V2),
(a1,V3), (a1,V4), (a2,V1), (a2,V2), (a2,V3), (a2,V4),]. The players of a
game are not required to have the same strategy set. The
preferences of the players in the game involve the expected
utilities of the strategy combinations. The players wish to maximize their expected utility received by playing the game.
6.2. Context of the game
It is important to understand how games are used in determining the egress alternative and velocity for the players. An important concept in game theory is the beliefs that players have on
other players. A game formation that captures all of the beliefs of
each player on the other players traits is a Bayesian game [41].
In this formation, each player has multiple types. Each type is a
single combination of individual traits such as most efcient
velocity, follow rate, hazard rate, and so on. The number of player
types increase as the number of possible traits and the number of
possible trait values (such as multiple believed most efcient
velocities) increase. The beliefs each player has on another player
are based on the populations trait distributions. These distributions are common knowledge, meaning that each player knows
the basis of every other players beliefs. Common knowledge is a
requirement in Bayesian games. Bayesian games are mathematically
complex due to each type of a player being viewed as a separate
player in a much larger game. As the number of players in a game
increases, the complexity of the game increases. While Bayesian
games enable the most accurate determination of each players
alternative decision, these mathematical complexities must be
understood. A trade-off is made in this paper to reduce computational complexity by sacricing game formation perfection.
When all of the players know all of the other players traits
(e.g. the other players follow rates, most efcient velocity, risk
preference, etc.), then a simultaneous decision is made using game
theory so that the resulting equilibrium is used to dictate all of the
players decisions. The situation investigated in this paper (and which
occurs in the real world) is that the players do not know everything
about the other players. This conicts with the game theory
requirement that the knowledge of the other individuals beliefs is
known. To overcome this, the games are set in the viewpoint of the
individual that is currently making the decision, POI.

127

Different games can be formed by the players in a decision

region to represent the game they believe they are playing. For
instance, two players in the same decision region may be playing
different games, as these are based on individual beliefs. These
games can have vastly different Nash equilibria (the optimum
choice set), where each equilibrium is believed to be true to the
related individual who forms the game (the POI). In the games
examined in this paper, it is assumed that the POI reects his
beliefs on to the other players. The POIs opponents are assumed to
have the same environmental beliefs and hazard rates as the POI.
For this paper, the POI assumes the other players risk preferences
and follow rates to be neutral.
Different game forms are used to depict the interactions of the
evacuees in order to reduce mathematical complexity while
maintaining high delity models of the most inuential interactions. The trade-offs for the different game forms are seen in
Table 1. The form of the game depends on the number of evacuees
playing the game. Bayesian games are played for the smallest
games of 2 players. A pseudo Bayesian game is played for slightly
larger that are less than 4 players. A simplied game is played for
very large games with 4 or more players. These different forms
have been developed in this work to maximize mathematical
accuracy for small games and to reduce accuracy in larger games
in order to improve computational efciency.
6.3. Bayesian games
Bayesian games are designed to represent players beliefs on
the other players in the game. In evacuations the players can have
many beliefs on the other players, such as their believed most
efcient velocity, follow rate, hazard rate and so on. Bayesian
games have not been used in any form to model evacuations in the
past, and hence are a novel concept. Evacuation simulations
benet from the incorporation of Bayesian games by enabling
the impact of beliefs on the decision making process and ultimately each individuals behaviour. For the Bayesian game studied
in this paper the beliefs of the POI on the other players most
efcient velocity are incorporated. This game is viewed from the
eyes of the POI. Our assumptions in this game is that the POI
believes that the opponents most efcient velocity probability
distribution is the distribution [29] related to the opponents
characteristics, such as age and sex. It is also assumed that the
opponent knows exactly what the POIs most efcient velocity is,
possibly due to careful observation. These beliefs of the POI and
the opponent are common knowledge, meaning that each player
knows the other players beliefs. Since the game is in the eyes of
the POI, he assumes that the opponent has some of the same
beliefs as him, such as hazard rate and environmental beliefs.
Player types are formed based on player beliefs. The player
types are based on most efcient velocities. Since the POIs most
efcient velocity is known, he has a singular type. The opponent
has a certain number of possible most efcient velocities, which
result in the same number of opponent types. The strategies of the
players are the combinations of the egress alternatives in the
decision region and the set of velocities to be investigated. The set
of velocities consists of the most efcient velocities of the POI, the
most efcient velocities of the types of the opponents and other

Table 1
Game form trade-offs.
Beliefs on opponents

Game form

Fidelity

Computational efciency

Singular beliefs
Probabilistic beliefs on some opponents
Probabilistic beliefs on all opponents

Simplied game (game theory)

Pseudo-Bayesian game (game theory and Bayesian game theory)
Bayesian game (Bayesian game theory)

Low
Medium
High

High
Medium
Low

128

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

velocities that are slower and faster than the most efcient
velocities. The more velocities that are examined the more
accurate the games result; however, the number of velocities
must be limited in order to maintain computational efciency.
In the game examples, 5 additional velocities are chosen at equal
distance from each other spanning the range of slightly above 0 to
the POIs maximum velocity. Each type of a player must have the
same strategy set; however, the players themselves do not have to
have the same strategy set.
Fig. 6 shows an example of the form of a game between two
players each of a single type. In Fig. 6, the strategies (set of
alternative and velocity) for the POI player are listed vertically on
the exterior of the game form and the opponents strategies are
listed horizontally. Inside the game form are the expected utilities
for each player associated with the strategy sets of the two players.
Each of these expected utility sets is found at the intersecting box
of the associated player strategies. The expected utilities for each
player may be impacted by the strategy chosen by the other player.
This is a standard visualization tool in game theory [42].
The full form of a Bayesian game where the non-POI player has
5 types is seen in Fig. 7. In Fig. 7 each combination of player types
(referred to as states) consists of the game form [43] seen in Fig. 6.
The set of strategies for each player is the same for the 5 different
states, as is consistent with Bayesian game theory. The opponents

possible most efcient velocities are determined using the distribution related to the opponents sex and age. In this particular
game form the states are related to percentiles correlating to the
opponents most efcient velocity distribution. Below each state
are each players types, as well as the probabilities each player
believes that the other player is of that type. For example, in state
B the POI believes that the opponent is 22% likely to be of type 2.
The probability of each of these states is a result of the use of a
triangular distribution to capture the POIs belief on the opponents most efcient velocity. The use of different beliefs will result
in different Bayes-Nash equilibriums. The triangular distribution
was chosen due to its nature of approximating normal distributions. The distribution used is fully dependent on the beliefs of the
POI, and is likely to not be a triangular distribution after updating
from new information. Since the POI is of one type the opponent
knows for certainty what type the POI is. The expected utilities
that ll the states are calculated using Eq. (7).
A simple example of a Bayesian game is seen in Fig. 8. In this
gure the POI has a belief on another individual, player 2. POI
believes that player 2 has two possible most efcient velocities,
slow and fast. The POI believes that player 2 has a slow most
efcient velocity with 70% certainty and a fast most efcient
velocity with 30% certainty. Player 2 knows the most efcient
velocity of the POI; hence the POI is of a single type in the Bayesian
game. Each player has two alternatives (POI with A and B, and
player 2 with C and D). Within the outcome boxes of the Bayesian
game are the utility values associated with the POI and player 2 (in
the upper left and lower right corners respectively). The blue
boxed utilities represent each of player 2s types best responses
given that the POI chooses a pure strategy of A or B. The POI views
the scenario as a 3 player game. The POIs viewpoint is seen in
Fig. 9, and is used to determine the POIs best responses to a set of
opponent actions. The outcome values in Fig. 9 represent the
expected utilities for the POI given the set of alternatives chosen

Fig. 6. Bayesian game state example.

Fig. 8. Bayesian game numerical example.

Fig. 7. Bayesian game example.

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

by himself and the two player 2 types. Also included in Fig. 9 are
the best responses of each of player 2s types represented by the
blue boxed S and F. From Fig. 9, the Bayes-Nash equilibrium can be
determined as [A, (C,D)]. Given this equilibrium the methodology
would determine the POIs decision to be A.
6.4. Pseudo-Bayesian games
As the number of players in a Bayesian game that have multiple
types increases, so does the complexity of solving the game. This
complexity is due to the consequence that each player type is
viewed as a separate entity in the game (a separate player).
In order to nd a balance between the ability to incorporate
beliefs on an opponent and the simplicity of making singular
assumptions on an opponent, a pseudo-Bayesian game is created.
A pseudo-Bayesian game is formed when the number of people
in a game is less than a certain number. For computational
efciency, the number used in this paper is 4 players. In the
pseudo game, Bayesian players are opponents in the pseudo game
whom the POI has beliefs on their most efcient velocities. Game
players are opponents whom the POI assumes singular beliefs on
their most efcient velocities. The pseudo game is a mixture of
players found in a Bayesian game and players found in a normal
game. The determination of which players are Bayesian and which

Fig. 9. POI numerical example. (For interpretation of the references to color in this
gure legend, the reader is referred to the web version of this article.)

Fig. 10. Game theory example scenario.

V(m/s),A
1.55, 0
.015, 0
2.41, 0
4.78, 0
1.55, 1
.015, 1
2.41, 1
4.78, 1

1.55, 0 .015, 0
-.852, -.852,
-.65
-8.5
-18.5, -18.5,
-.65
-.580
-.88,
-.88,
-.65
-8.5
-.93,
-.93,
-.65
-8.5
-.65,
-.65,
-.584
-8.5
-8.5,
-8.5,
-.65
-8.5
-.66,
-.66,
-.65
-8.5
-.68,
-.68,
-.65
-8.5

129

are Game is related to the distance of each to the POI. It is assumed

that the closer a player is to the POI, the more likely that player is
competing to reach an egress alternative rst and therefore will
affect the POIs expected utility more. Therefore, the closer the
player is, the more likely the POIs expected utility is affected by
that players decisions and the more accurately that person should
be modelled. The pseudo-game is still a Bayesian game and hence
the Game players are assumed to have the same beliefs as the POI
concerning the Bayesian players. The Game players most efcient
velocities are assumed to be the same percentile in their probability distribution as the POIs most efcient velocity percentile,
dependent on their age and sex. For example, if the POI is of the
43rd percentile for most efcient velocity in his population group
then he would assume the Game players are of the 43rd percentile
for most efcient velocity in their population group.

6.5. Simplied games

As the number of players in the pseudo-Bayesian game
increases, the complexity of solving the game increases as well.
A simplied game is used in order to ease the computational
requirements of games with many players. The simplied game is
a game formed of Game players and of Assumed players. Game
players are the same as seen in the pseudo-Bayesian game, with
their believed most efcient velocity based off of the POIs
percentile. Assumed players only have one strategy in the game,
hence their strategy is pre-chosen. An Assumed players strategy is
the same as their last chosen strategy (in terms of egress alternative and velocity) as observed by the POI. For example, a POI is
in a decision region with 10 people. The POI views the farthest
opponent as an Assumed player. If the Assumed player is moving
with a certain velocity towards a specic egress alternative then
the POI assumes the player uses this strategy in the simplied
game. In this game form, the Assumed players are no longer
viewed as interacting opponents; they are viewed as obstacles
with predetermined movements. The Game players are viewed as
interacting opponents and are the players closest to the POI. The
number of Game players depends on the accuracy desired. In this
paper, a limit of 3 Game players is used, with the rest of the
individuals in a decision region being Assumed players. If the POI
has not observed an Assumed players previous strategy, then
that player is assumed to be stationary for that games duration.
This will occur in the case where the POI has just entered a
decision region or at simulation start.

2.41, 0 4.78, 0
-.852, -.852,
-.66
-.68
-18.5, -18.5,
-.66
-.68
-.88,
-.88,
-.66
-.68
-.93,
-.93,
-.66
-.68
-.65,
-.65,
-.66
-.68
-8.5,
-8.5,
-.66
-.68
-.66,
-.66,
-.66
-.68
-.68,
-.68,
-.66
-.68

1.55, 1 .015, 1
-.852, -.852,
-.852
-18.5
-18.5, -18.5,
-.852
-18.5
-.88,
-.88,
-.852
-18.5
-.93,
-.93,
-.852
-18.5
-.65,
-.65,
-.852
-18.5
-8.5,
-8.5,
-.852
-18.5
-.66,
-.66,
-.852
-18.5
-.68,
-.68,
-.852
-18.5

Fig. 11. Game theory example normal form.

2.41, 1
-.852,
-.88
-18.5,
-.88
-.88,
-.88
-.93,
-.88
-.65,
-.88
-8.5,
-.88
-.66,
-.88
-.68,
-.88

4.78, 1
-.852,
-.93
-18.5,
-.93
-.88,
-.93
-.93,
-.93
-.65,
-.93
-8.5,
-.93
-.66,
-.93
-.68,
-.93

130

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

V(m/s),A
1.55, 0
.015, 0
2.41, 0
4.78, 0
1.55, 1
.015, 1
2.41, 1
4.78, 1

1.55, 0
-.35,
-.65
-18,
-.65
-.38,
-.65
-.43,
-.65
-.65,
-.584
-8.5,
-.65
-.66,
-.65
-.68,
-.65

.015, 0
-.35,
-8.5
-18,
-.580
-.38,
-8.5
-.43,
-8.5
-.65,
-8.5
-8.5,
-8.5
-.66,
-8.5
-.68,
-8.5

2.41, 0
-.35,
-.66
-18,
-.66
-.38,
-.66
-.43,
-.66
-.65,
-.66
-8.5,
-.66
-.66,
-.66
-.68,
-.66

4.78, 0
-.35,
-.68
-18,
-.68
-.38,
-.68
-.43,
-.68
-.65,
-.68
-8.5,
-.68
-.66,
-.68
-.68,
-.68

1.55, 1
-.852,
-.852
-18.5,
-.852
-.88,
-.852
-.93,
-.852
-.15,
-.852
-7.7,
-.852
-.16,
-.852
-.18,
-.852

.015, 1
-.852,
-18.5
-18.5,
-18.5
-.88,
-18.5
-.93,
-18.5
-.15,
-18.5
-7.7,
-18.5
-.16,
-18.5
-.18,
-18.5

2.41, 1
-.852,
-.88
-18.5,
-.88
-.88,
-.88
-.93,
-.88
-.15,
-.88
-7.7,
-.88
-.16,
-.88
-.18,
-.88

4.78, 1
-.852,
-.93
-18.5,
-.93
-.88,
-.93
-.93,
-.93
-.15,
-.93
-7.7,
-.93
-.16,
-.93
-.18,
-.93

Fig. 12. Game theory example normal form individual wants to follow.

6.6. Examples
In this section four examples are presented in order to clearly
represent evacuee interactions as well as the effects of following
others and waiting in queues.
6.6.1. Example 1: Following othersNeutral
The rst example displays the effects of a player with and
without a strong desire to follow another player through an
egress alternative. The environment and initial player positions
are seen in Fig. 10. In this example, egress alternatives 0 and 1 are
identical, with the exception of location. We will rst examine
the situation where both players, whom are identical in traits
(identical to the individual in Section 5.4), are neutral in terms of
their follow rates. In this example Player A is assumed to be a
Game player rather than a Bayesian player. In this example player
X is the POI with a neutral following preference. The game played
from POIs point of view is seen in Fig. 11. The strategies for player
X are seen in the rst column, (velocity, alternative), and player
As strategies are seen in the rst row. The outcome cells are
represented by the boxes at the intersection of strategies. The
outcome cells show the expected utilities for the players given
that strategy combination, with player Xs expected utility
followed by player As. The grey shaded squares indicate Player
Xs preferred strategy to each of Player As strategies. The green
shaded squares indicate Player As preferred strategy to each of
Player Xs strategies. The games solution (Nash equilibrium [42])
is seen highlighted in red. Player As 1st strategy strictly dominates all of his other strategies, as does player Xs 5th strategy.
The equilibrium for the game that Player X believes he is playing
is a pure strategy Nash equilibrium. Player X believes that player
A will choose (1.55 m/s, alternative 0) and he himself should
choose (1.55 m/s, alternative 1). This equilibrium results in each
player choosing their closest door and moving at their most
efcient velocities.

Fig. 13. Game theory example scenario one exit.

Fig. 14. Game theory example many people.

6.6.2. Example 2: Following othersNon-neutral

The same situation is examined but player X now holds a nonzero follow rate. Player X now has a following parameter of  0.5.
A negative following parameter means that he deems it very
worthwhile to follow people, as he believes he gains half of a
calorie by following someone. Player X still believes Player A has a
neutral follow rate. The game that player X believes he is playing is
seen in Fig. 12. In Fig. 12, it is shown that player A continues using

his strictly dominant strategy. Player X chooses a strategy which

coordinates with the alternative strategy of player A, due to his
increased following rate. The games Nash equilibrium is when
player A chooses (1.55 m/s, alternative 0) and player X chooses
(1.55 m/s, alternative 0). In this equilibrium the POI is choosing the
alternative that is farther away, but is worth more to him due to
the desire of following player A.

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

131

Table 2
Corridor evacuation experimental results normal evacuation.
Objective: evacuate normally

Total evacuation time (s)

Experiment Trial 1
Experiment Trial 2

35
32

Table 3
Corridor evacuation experimental results min individual evacuation time.
Objective: minimize
individual EVAC time

Total evacuation time (s)

Experiment
Experiment
Experiment
Experiment

18
19
20
19

Trial
Trial
Trial
Trial

1
2
3
4

Fig. 15. Corridor evacuation environment.

simulation were realistic and determined by decision processes

captured through the methods described in the paper.

7. Validation
6.6.3. Example 3: Waiting in queue
The next scenario investigates the effects waiting time has on
the velocity chosen by a player. The environment and initial player
positions for this scenario are seen in Fig. 13. The players traits are
identical. There exists a single egress alternative for this environment. If player X, the POI, were in the room alone then the velocity
that he would choose would be his most efcient velocity of
1.55 m/s. When another player is in the room the POIs utility is
now dependent on the velocity chosen by the other player. The
utility is affected as the POIs wait time at the alternative is
impacted. This wait time is due to the opponent possibly reaching
the alternative slightly earlier than the POI. Each player, in their
respective games that they believe they are playing, has a solution
that is a mixed strategy Nash equilibrium [42]. The strategies with
non-zero probabilities in the mixed Nash equilibrium are for the
individuals to move at their most efcient speed or move at a
slightly faster speed. A velocity that is slightly faster than an
individuals most efcient velocity causes more energy being
consumed moving but less energy consumed waiting if they beat
the opponent to the exit. Since the equilibrium is a mixed strategy,
each simulation run may result in a different strategy being chosen
by each player. This is due to mixed strategies relating to multiple
strategy sets having a probability of being selected. Through this
competition to reduce their waiting times the players velocities
are affected.

6.6.4. Example 4: Large number of players

For the nal example a larger environment with multiple
individuals with randomly assigned traits is created. This environment is seen in Fig. 14. The simulation is performed in order to
ensure that when traits are randomized, non-realistic behaviours
do not emerge. The simulation shows that 7 people initially chose
the left alternative and 8 people initially chose the right. One
individual near the left alternative but at the end of the queue
chose the right alternative, resulting in a collision with another
individual. A primitive collision function is used to cause individuals to stop if their velocity causes a movement through another
individual. On the next decision time (for this simulation the time
steps are in single seconds) fewer people are in line for the left
alternative, causing the colliding evacuees to both choose the
closest alternative, being the left. Competition to reduce waiting
time is seen to take place between individuals in the environment
within the smaller top corridors. This competition is sparked due
to the top alternatives having a low ow rate caused by the
alternatives short widths. All movements examined in the

The previous section focused on simple examples to illustrate

that the methodology in this paper captures human tendencies
and characteristics in exit selection. In this section, multiple case
studies are presented that are used to validate the Vacate-GT
emergency egress simulator. One of the case studies is discussed in
detail that pertains specically to exit selection.
7.1. Corridor evacuation validation
In 2010, an evacuation experiment [14] was performed that
focused on the exit decisions and evacuation times of a group of
individuals in a corridor. In this experiment 54 people evacuated
from a small room into a much larger room through one of two
exits situated along a hallway. The population was composed of
college undergraduates. Different objectives were given to the
population for different trial runs. These objectives were to
evacuate normally, evacuate minimizing individual time and
evacuate minimizing the group time. For a 54 person population,
the objective of evacuating normally was performed in two trials
and the other objectives were performed in four trials each. The
results of the corridor evacuations are approximated from the
charts in [14] and seen in Tables 2 and 3. The average percentage
of the population choosing the closer exit in the experiments with
the objective of minimizing individual evacuation time is 67.59%.
This experiment is especially useful to the methodology introduced in this paper as it explores both evacuation time and exit
selection.
7.1.1. Simulation setup
Vacate-GT is validated by simulating two of the objectives
given to the population in the trials. These objectives are to
evacuate normally and to evacuate minimizing individual time.
Each objective is simulated 3 times with randomly determined
traits of the evacuees. The only requirement imposed on the traits
is that the evacuees be between the ages of 18 and 22 in order to
simulate college undergraduates. The evacuation time is determined by the last person exiting the corridor. The percentage of
people choosing the closer exit is also calculated. An example of
how identical populations can result in different evacuations due
to randomization of the selection of the Nash equilibriums is also
examined using the experimental setup.
The geometry of the experiment [14] is replicated in the
validation simulations with the addition of a large exit on the far
side of the large room. Fig. 15 displays the simulation environment
with the initial positions of the evacuees. The closest exit is seen as
the further left of the two corridor exits, drawn in dark blue along

132

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

Fig. 16. Corridor evacuation identical population different decisions.

Table 4
Corridor evacuation identical population comparison.
Objective evacuate normally
identical population

Evacuation
time (s)

(%) of Population choosing

closest exit

Simulation Norm 11
Simulation Norm 12

33.1
32.8

61
66

Table 5
Corridor evacuation simulation results normal evacuation.
Objective evacuate normally
3 random populations

Evacuation
time (s)

(%) of Population
choosing closest exit

Simulation Norm 11
Simulation Norm 2
Simulation Norm 3

33.1
28.5
29.5

61
55
59

Fig. 18. Corridor evacuation results min individual time.

Table 7
Other case study validations.

Fig. 17. Corridor evacuation results normal evacuation.

Table 6
Corridor evacuation simulation results min individual evacuation time.
Objective min individual. Time
3 random populations

Evacuation
time (s)

(%) of Population
choosing closest exit

Simulation Min Ind 1

Simulation Min Ind 2
Simulation Min Ind 3

17
19.2
16.8

61
63
66

Study

Experimental
evacuation time (s)

Average (3 runs) simulated

evacuation time (s)

Stapelfeldt (0.8m)
Stapelfeldt (1.5m)
B737-200

50
30
30.5

51.3
34.9
29.4

the hallway. The large room after the hallway doors allows for the
inuence of the people just entering the large room to be
simulated on the people in the process of entering.
For the evacuate normally objective, the hazard rates of the
individuals are randomly chosen amongst a range of relatively low
rates to imply the low hazard of the region. For the minimize
individual time objective, the hazard rates for the population are
set to a constant high value of 0.06 cal/s.

7.1.2. Results
Two simulations are performed with identical populations to
illustrate the possibility of different evacuations occurring. These
simulations consist of evacuees with identical traits. The objectives
of the simulations are to evacuate normally. Fig. 16 shows
visualizations of the two simulations after ten seconds from
simulation start. As can be seen, the evacuees have chosen
different velocities and possibly different egress alternatives that

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

have resulted in different positions. Numerically the simulations

can be seen to be different from Table 4. This table displays the
differences in evacuation time and egress alternative selection
between the two simulation populations. These variations in
performance are due to the inherently random nature of Nash
equilibrium selection when multiple equilibriums are present as
well as the random nature of mixed Nash equilibriums.
Three simulations using the methodology discussed previously
are performed that capture the populations characteristics. Population characteristics that are not described in Ref. [14], are
randomized, such as gender positioning, participant height, and
most efcient velocities. The simulated data for the evacuate
normally objective is seen in Table 5. A comparison between
experimental and simulation data is displayed visually in Fig. 17.
The average simulated evacuation time is 30.36 s, which is 10%
error of the experimental average.
In order to simulate the scenario in which the individuals
objectives are to minimize their evacuation time, a sense of
urgency is created through a high hazard rate of 0.06 cal/s. This
energy consumption rate causes the individuals to prefer to move
quickly in order to reduce the amount of time spent in the
environment. The simulated data is seen in Table 6 and compared
visually with the experimental data in Fig. 18. The average
simulated evacuation time is 17.66 s, which is 7% error of the
experimental average. The average percentage of population
choosing the closer of the two exits relative to the starting
positions for the simulated data is 63.33%, which is 6.3% error of
the experimental data. The error is due to many variables, such as
population characteristics, a low number of experimental data
points, the amount of urgency addressed to the participants, and
assumptions in the methodology.
7.2. Further validation case studies
The corridor evacuation case study is unique in that it provides
a source of validation for both movement and exit selection
decisions. Vacate-GT is also validated against two other experiments, a Boeing 737-200 evacuation [13] and a single room
evacuation (Stapelfeldt trials) [12], focusing on movement decisions. These additional case studies do not focus on door selection
and were only used to validate the evacuation time performance of
Vacate-GT. The results of the experiments and simulations are
displayed in Table 7. The results show that Vacate-GT can
accurately capture different evacuation events through the use of
decision analysis.

8. Summary
Human tendencies in evacuations have historically been difcult to model in simulations. Traditionally, evacuation simulators
have used heuristics or primitive objective functions in order to
capture crowd ow. Many of the methods that have been previously proposed have not included the interactions between
evacuees during the egress alternative decision process. As seen
in this paper, the mathematical frameworks of game theory and
decision theory provide a high delity model for describing the
exit decision making process numerically. The decision making
process determines the behaviours of the evacuees, eliminating
the need to model the behaviours directly. A novel formation of
evacuation games was created where the players were the
individual evacuees. Multiple game forms with varying delities
were created to enable trade-offs between accuracy and computational complexity. A unique utility function based on energy
expenditure was developed for exit alternative selection. Furthermore, the incorporation of the evacuees beliefs on the

133

environment (updated using Bayes theorem) and their effects on

the decision making processes were examined.
In this paper, multiple examples show that evacuee tendencies,
such as risk preferences, velocity determination, and following
others, can be simulated using decision analysis. These examples
have shown how the egress alternative decision making process
can be handled using proven mathematical methods to capture
individual decisions. Case studies demonstrated the validity of
Vacate-GT compared to experimental data. A method using
decision analysis tools that examines individual decision making
offers a high delity alternative to traditional large-scale heuristic
methods.

Acknowledgements
This research was funded in part by the National Science
Foundation. The opinions, ndings, and conclusions stated herein
are those of the authors and do not necessarily reect those of the
sponsor.
References
[1] S. Tzu-Sheng, ESM: a building evacuation simulation model, Build. Environ. 40 (5)
(2005) 671680.
[2] M. Owen, E.R. Galea, P.J. Lawrence, The exodus evacuation model applied to
building evacuation scenarios, J. Fire. Prot. Eng. 8 (2) (1996) 6584.
[3] S.M. Lo, et al., An evacuation model: the SGEM package, Fire Saf. J. 39 (3)
(2004) 169190.
[4] P.A. Thompson, E.W. Marchant, A computer model for the evacuation of large
building populations, Fire Saf. J. 24 (2) (1995) 131148.
[5] B.L. Mesmer, C.L. Bloebaum, Importance of incorporation of personal communication devices in evacuation simulators, Saf. Sci. 50 (5) (2012) 13131318.
[6] Z. Xue, A Particle Swarm Optimization Based Multi-agent Stochastic Evacuation Simulation Model, State University of New York at Buffalo, United States
New York, NY, 2009.
[7] B. Ries, A Heuristic Optimization Based Emergency Evacuation Simulation
Approach for a Building and its Surrounding Environment State University of
New York at Buffalo, United StatesNew York, NY (2006) 136 (p).
[8] B.L., Mesmer , C.L. Bloebaum, Modeling decision and game theory based pedestrian
velocity vector decision in emergency evacuations, in: Second International
Conference on Evacuation Modeling and Management 2012: Chicago, Illinois, USA.
[9] D. Bernoulli, Exposition of a new theory on the measurement of risk,
Econometrica 22 (1) (1954) 2336.
[10] J. Von Neumann, O. Morgenstern, Theory of Games and Economic Behavior,
Princeton University Press, Princeton, 1944.
[11] P. Fiorini, Z. Shiller, Motion planning in dynamic environments using velocity
obstacles, Int. J. Rob. Res. 17 (7) (1998) 760772.
[12] J.P. Stapelfeldt, Angst- und Paknikstande aus der Sicht des Brandschutzes
(Conditions of anxiety and panic from the viewpoint of re protection), vfdbZ. (1986) 2.
[13] H. Muir, A. Cobbett, Aircraft Evacuation TestsAn Initial Assessment of the
Inuence of Various Aisle Congurations and Lighting Conditions Under
Different Evacuation Scenarios, T.D.C.S.a.S.T. Canada, (Ed.) 1996.
[14] S. Helio
vaara, et al., Pedestrian behavior and exit selection in evacuation of a
corridoran experimental study, Saf. Sci. 50 (2) (2012) 221227.
[15] J. v. Neumann, Zur Theorie der Gesellschaftsspiele, Math. Ann. 100 (1) (1928)
295320.
[16] S. Lo, et al., A game theory based exit selection model for evacuation, Fire Saf. J.
41 (5) (2006) 364369.
[17] S. Hoogendoorn, P.H.L. Bovy, Simulation of pedestrian ows by optimal control
and differential games, Optimal Control Appl. Methods 24 (3) (2003) 153172.
[18] C. Dogb, Modeling crowd dynamics by the mean-eld limit approach, Math.
Comput. Modell. 52 (910) (2010) 15061520.
[19] J. Tanimoto, A. Hagishima, Y. Tanaka, Study of bottleneck effect at an
emergency evacuation exit using cellular automata model, mean eld approximation analysis, and game theory, Physica A 389 (24) (2010) 56115618.
[20] X. Zheng, Y. Cheng, Conict game in evacuation process: a study combining
Cellular Automata model, Physica A 390 (6) (2011) 10421050.
[21] X. Zheng, Y. Cheng, Modeling cooperative and competitive behaviors in
emergency evacuation: a game-theoretical approach, Comput. Math. Appl.
62 (12) (2011) 46274634.
[22] G.A. Cavagna, M. Kaneko, Mechanical work and efciency in level walking and
running, J. Physiol. 268 (2) (1977) 467481.
[23] G.A. Cavagna, P.A. Willems, N.C. Heglund, The role of gravity in human
walking: pendular energy exchange, external work and optimal speed,
J. Physiol. 528 (3) (2000) 657668.
[24] B. Givoni, R.F. Goldman, Predicting metabolic energy cost, J. Appl. Physiol. 30
(3) (1971) 429433.

134

B.L. Mesmer, C.L. Bloebaum / Fire Safety Journal 67 (2014) 121134

[25] K.B. Pandolf, B. Givoni, R.F. Goldman, Predicting energy expenditure with loads
while standing or walking very slowly, J. Appl. Physiol. 43 (4) (1977) 577581.
[26] H.J. Ralston, Energy-speed relation and optimal speed during level walking,
Int. Z. Angew. Physiol. 17 (4) (1958) 277283.
[27] H.J. Ralston, in: R.M. Herman, S. Grillner, P.S.G. Stein, D.G. Stuart (Eds.),
Energetics of Human Walking, in Neural Control of Locomotion, Plenum Press,
New York, 1976, pp. 7798.
[28] W.H. Van der Walt, C.H. Wyndham, An equation for prediction of energy
expenditure of walking and running, J. Appl. Physiol. 34 (5) (1973) 559563.
[29] R. Knoblauch, M. Pietrucha, M. Nitzburg, Field studies of pedestrian walking
speed and start-up time, Transp. Res. Rec. J. Trans. Res. Board 1538 (1996)
2738.
[30] W.T. Scherer, T.A. Pomroy, D.N. Fuller, The triangular density to approximate
the normal density: decision rules-of-thumb, Reliabil. Eng. Syst. Saf. 82 (3)
(2003) 331341.
[31] H.W.,Stoudt , Skinfolds, Body Girths, Biacromial Diameter, and Selected
Anthropometric Indices of Adults, United States, 19601962. National Center
for Health Statistics Vital and Health Statistics Series 11: Data from the
National Health Survey Data from the Health Examination Survey no 35.
1970, Washington,: U.S. Health Services and Mental Health Administration. iv,
63 p.
[32] C.L. Ogden, et al., Mean body weight, height, and body mass index, United
States 19602002, Adv. Data 347 (2004) 117.
[33] Anthropometric Reference Data for Children and Adults United States, 2003
2006. National Health Statistics Reports. (Ed.) M.A. McDowell. 2008, U.S. Dept.

[34]
[35]

[36]

[37]
[38]
[39]

[40]
[41]
[42]
[43]

of Health and Human Services Centers for Disease Control and Prevention
National Center for Health Statistics: Hyattsville, MD.
S.M.V. Gwynne, et al., Questioning the linear relationship between doorway
width and achievable ow rate, Fire Saf. J. 44 (1) (2009) 8087.
M. Bayes, M. Price, An Essay towards Solving a Problem in the Doctrine of
Chances. By the Late Rev. Mr. Bayes, F. R. S. Communicated by Mr. Price, in a
Letter to John Canton, A. M. F. R. S. Philosophical Transactions (16831775),
1763. 53: pp. 370418.
S. Alpanda, G. Woglom, The Case Against Power Utility and a Suggested
Alternative: Resurrecting Exponential Utility, University Library of Munich,
Germany, 2007.
E.W. Dijkstra, A note on two problems in connexion with graphs, Numerische
Math. 1 (1) (1959) 269271.
R. Waldura Dijkstras Shortest Path Algorithm in Java. 2007 8-20-2007 [cited
2012 3-28]; Available from: http://renaud.waldura.com/doc/java/dijkstra/.
J. Kennedy, R. Eberhart, Particle swarm optimization. in Neural Networks,
1995, in: Proceedings., IEEE International Conference on. 1995. Perth, WA,
Australia: IEEE.
J.F. Nash, Equilibrium Points in N-Person Games, Proc. Nat. Acad. Sci. U.S.A. 36
(1) (1950) 4849.
J.C. Harsanyi, Games with Incomplete Information, Am. Econ. Rev. 85 (3)
(1995) 291303.
K.G. Binmore, Playing for Real: A Text on Game Theory, ix, Oxford University
Press, Oxford; New York (2007) 639.
M.J. Osborne, An Introduction to Game Theory, xvii, Oxford University Press,
New York, NY (2004) 533.